GPU workshop cheatsheet, Slides of Data Structures and Algorithms

Download GPU workshop cheatsheet and more Slides Data Structures and Algorithms in PDF only on Docsity! 1 GPU workshop cheatsheet OpenACC guide (PGI flags –acc –ta=nvidia –Minfo=accel) Kernels Construct An accelerator kernels construct surrounds loops to be executed on the accelerator, typically as a sequence of kernel operations. C #pragma acc kernels [clause [[,] clause]…] new-line { structured block } Fortran !$acc kernels [clause [[,] clause]…] structured block !$acc end kernels Any data clause is allowed. other clauses if( condition ) When the condition is nonzero or .TRUE. the kernels region will execute on the accelerator; otherwise, it will execute on the host. async( expression ) The kernels region executes asynchronously with the host. Data Construct An accelerator data construct defines a region of the program within which data is accessible by the accelerator. C #pragma acc data [clause[[,] clause]…] new-line { structured block } Fortran !$acc data [clause[[,] clause]…] structured block !$acc end data Any data Data Clauses The description applies to the clauses used on parallel constructs, kernels constructs, data constructs, declare constructs, and update directives. copy( list ) Allocates the data in list on the accelerator and copies the data from the host to the accelerator when entering the region, and copies the data from the accelerator to the host when exiting the region. copyin( list ) Allocates the data in list on the accelerator and copies the data from the host to the accelerator when entering the region. copyout( list ) Allocates the data in list on the accelerator and copies the data from the accelerator to the host when exiting the region. create( list ) Allocates the data in list on the accelerator, but does not copy data between the host and device. present( list ) The data in list must be already present on the accelerator, from some containing data region; that accelerator copy is found and used. 2 CUDA Built-in kernel variables  gridDim.[x,y,z] -> Three dimensional vector containing the dimensions of the grid. This is a constant that is set at kernel launch time. If not set explicitly each dimension defaults to 1.  blockIdx.[x,y,z] -> Three dimensional vector containing the block index within the grid. This is a dynamic value that depends on which block calls it.  blockDim.[x,y,z] -> Three dimensional vector containing the dimensions of the thread block. This is set at kernel launch time. If not set explicitly each dimension defaults to 1.  threadIdx.[x,y,z] -> Three dimensional vector specifying the thread index within the thread block. Dynamic value depending on which thread calls it. Important Functions  Kernel Launch o Kernel_name<<< gridsize, blocksize >>>(arg1,arg2,…);  Memory Management o cudaError_t cudaMalloc( void **devPtr, size_t size );  Example: cudaMalloc( (void **) &d_c, numbytes ); o cudaError_t cudaFree( void *devPtr );  Example: cudaFree( d_c ); o cudaError_t cudaMemcpy( void *dst, const void *src, size_t size, enum cudaMemcpyKind kind );  enum cudaMemcpyKind  cudaMemcpyHostToDevice  cudaMemcpyDeviceToHost  cudaMemcpyDeviceToDevice  Example: cudaMemcpy( d_c, c, numbytes,cudaMemcpyHostToDevice);  Error Checking o cudaError_t cudaGetLastError(void); o char* cudaGetErrorString( cudaError_t code ); o printf(“%s\n”, cudaGetErrorString( cudaGetLastError() ) );